Method forward

networks/psp_encoder_arch.py:129–250

(self, x, locs)

        

    def forward(self, x, locs):
        w_b = []
        extend_W = 32*4
        max_lr_width = x.size(3)
        for b in range(locs.size(0)): #locs: 0~2048
            x_for_w = []
            for c in range(locs.size(1)):
                center_loc = (locs[b][c]/4).int()
                start_x = max(0, center_loc - extend_W//2)
                end_x = min(center_loc + extend_W//2, max_lr_width)
                crop_x = x[b:b+1, :, :, start_x:end_x].detach()
                crop_x = self._check_outliers_pad(crop_x, start_x, end_x, max_lr_width, center_loc, extend_W) # 

                x_for_w.append(crop_x)
                # crop_x[...,62:66] = 1
                # save_image((crop_x+1)/2, 'trs_{}.png'.format(c))

            x_for_w = torch.cat(x_for_w, dim=0)
            
            x_c1 = self.conv1(x_for_w) #1
            x_c1 = self.relu(x_c1) 
            x_l1 = self.layer1(x_c1) #2
            x_l2 = self.layer2(x_l1) #1 [2, 64, 16, 256])
            x_l3 = self.layer3(x_l2) #2 torch.Size([2, 128, 8, 128]
            x_l4 = self.layer4(x_l3) #1 torch.Size([2, 256, 8, 128])
            x_l5 = self.layer5(x_l4) #2, torch.Size([2, 512, 4, 64])
            pyramid_x1 = _upsample_add(x_l5, self.layer256_to_512(x_l4))
            pyramid_x = self.layer512_to_outdim(pyramid_x1)
            w_each_b = self.feature2w(pyramid_x.view(pyramid_x.size(0), -1)) # 
            
            w_c = w_each_b
            w_b.append(w_c)
        w_b = torch.stack(w_b, dim=0)

        return w_b
 

 
        # w_b = []
        # for b in range(locs.size(0)): #locs: 0~2048
        #     w_c = []
        #     for c in range(locs.size(1)):
        #         if locs[b][c] < 2048:
        #             center_loc = (locs[b][c]/4).int() # 32*512
        #             start_x = center_loc - 16
        #             end_x = center_loc + 16
                    
        #             crop_x0 = x[b:b+1, :, :, start_x:end_x].clone()
        #             crop_x = self._check_outliers_pad(crop_x0, start_x, end_x) # 1, 512, 4, 4 or 1, 512, 8, 8
                    
        #             # save_image(crop_x[0], 'ss_{}.png'.format(c))
        #             x_c1 = self.conv1(crop_x) #1
        #             x_c1 = self.relu(x_c1) 
        #             x_l1 = self.layer1(x_c1) #2
        #             x_l2 = self.layer2(x_l1) #1 [2, 64, 16, 256])
        #             x_l3 = self.layer3(x_l2) #2 torch.Size([2, 128, 8, 128]
        #             x_l4 = self.layer4(x_l3) #1 torch.Size([2, 256, 8, 128])
        #             x_l5 = self.layer5(x_l4) #2, torch.Size([2, 512, 4, 64])